A small batch multi-variety precision strip steel statistical process control method

Abstract

The present application belongs to the technical field of precision strip steel production quality control, and particularly relates to a precision strip steel statistical process control method for small batch and multi-variety, and the specific steps are as follows: step S1: establishing a precision strip steel variety feature database; step S2: grouping historical varieties based on a K-means clustering algorithm; step S3: determining the cluster to which the current variety belongs; step S4: extracting historical quality data of the same cluster variety and removing outliers by using a 3σ criterion; step S5: data fusion based on similarity weighting; step S6: adaptive control limit calculation; step S7: multi-quality characteristic comprehensive control; step S8: abnormal mode identification; step S9: process capability evaluation; and step S10: control chart dynamic updating. The method effectively solves the problems of insufficient data, low historical data utilization, single-variable control limitations, fixed control limits and weak abnormal diagnosis capability by using K-means clustering analysis and data fusion technology and adaptive control limit calculation.

Description

Technical Field

[0001] This invention belongs to the field of precision strip steel production quality control technology, and in particular relates to a statistical process control method for precision strip steel with small batch and multiple varieties. Background Technology

[0002] Precision strip steel production is characterized by diverse product varieties, small-batch production, stringent quality requirements, and complex processes. The traditional Shewhart control chart method has the following problems in precision strip steel production: (1) Insufficient data volume Establishing a Shewhart control chart requires 25-30 sample data points, while small-batch varieties typically only have 10-20 samples per batch. This results in approximately 60-70% of varieties failing to establish an effective control chart, leading to inaccurate control limit calculations, a false positive rate as high as 8-15%, and a false negative rate of 15-25%. (2) Low utilization rate of historical data Existing methods only use historical data of the current variety. For new varieties or low-frequency varieties, the historical data accumulation period is long (6-12 months), the data utilization rate is less than 30%, and a large amount of data of similar varieties is not utilized. (3) Limitations of univariate control Traditional methods use univariate control charts, requiring the creation of more than 10 independent control charts. The correlation between quality characteristics is not considered, making comprehensive judgment difficult, and the accuracy of anomaly diagnosis is only 50-60%. (4) The control limits remain fixed. Once a traditional control chart is established, the control limits remain unchanged, making it unable to adapt to process improvement and optimization. In the early stages of new products, the control limits may be too wide or too narrow, and the control chart update cycle is long (usually 3-6 months). (5) Weak ability to diagnose abnormalities Existing methods can only determine whether something is abnormal, but cannot identify the type of abnormality. The analysis of the cause of abnormality relies on human experience, and there is no predictive or early warning function or effective abnormality diagnosis mechanism. Summary of the Invention

[0003] The purpose of this invention is to provide a statistical process control method for precision strip steel with small batch and multiple varieties, thereby overcoming the shortcomings in the aforementioned background technology.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A statistical process control method for precision strip steel with small batch and multiple varieties, the specific steps of which are as follows: Step S1: Establish a database of precision strip steel product characteristics Extract multidimensional feature vectors and perform Z-score standardization; Step S2: Group historical varieties based on K-means clustering algorithm The elbow rule is used to determine the optimal number of clusters, and the silhouette coefficient is used to evaluate the cluster quality. Step S3: Determine the cluster to which the current variety belongs. Calculate the Euclidean distance of the current variety to each cluster center, and assign it to the cluster with the closest distance; Step S4: Extract historical quality data of homo-clustered varieties and remove outliers using the 3σ criterion. Step S5: Similarity-weighted data fusion Variety similarity is calculated based on Euclidean distance, and data fusion is performed using similarity weighting to calculate fusion statistics; Step S6: Adaptive control limit calculation The weights of fused data and current variety data are dynamically adjusted as samples accumulate. Step S7: Integrated Control of Multiple Quality Characteristics The Hotelling T² statistic is used to achieve comprehensive monitoring of multiple quality characteristics, and the main sources of anomalies are identified through contribution analysis. Step S8: Abnormal Pattern Recognition Establish an anomaly pattern recognition mechanism and association rule base to achieve intelligent diagnosis of anomaly causes; Step S9: Process Capability Assessment Calculate the comprehensive process capability index of multiple quality characteristics to achieve accurate assessment of process capability; Step S10: Dynamically update the control chart Establish a dynamic update mechanism for control charts, including periodic updates of clustering, adaptive adjustment of fusion weights, and quality prediction and early warning.

[0005] Preferably, the specific content of step S1, establishing the precision strip steel product characteristic database, is as follows: (1) Define the variety feature vector For each precision strip steel grade, feature parameters are extracted to form a grade feature vector X: X=[x1,x2,x3,x4,x5,x6,x7,x8,x9,x 10 x 11 x 12[The following is a list of parameters:] where: x1 is thickness, ranging from 0.02 to 0.5 mm; x2 is width, ranging from 300 to 650 mm; x3 is material type (AISI 200, 300, 400 series); x4 is tensile strength, ranging from 300 to 2100 MPa; x5 is yield strength, ranging from 170 to 2000 MPa; x6 is elongation, ranging from 2% to 65%; x7 is surface roughness, Ra value 0.1 to 0.5 μm; x8 is thickness tolerance, ranging from ±0.001 to ±0.01 mm; x9 is hardness, ranging from 150 to 605 HV; 10 The number of rolling passes, ranging from 1 to 4; x 11 Annealing process type (bright annealing, bell-type furnace annealing, Ta annealing); x 12 For product series (soft state, hard state, extremely hard, Ta, special surface, ultra-thin foil); (2) Feature standardization Z-score normalization is applied to the variety feature vector: x' i =(x i -μ i ) / σ i , where μ i and σ i and are the mean and standard deviation of the i-th feature across all varieties, respectively.

[0006] Preferably, the specific details of grouping historical varieties in step S2 are as follows: (1) Determine the number of clusters The Elbow Method is used to determine the optimal number of clusters K, and the within-group sum of squares SSE corresponding to different K values ​​(K=3,4,5,...,15) is calculated as follows: SSE(K)=Σ i=1 K Σ X ∈C i ||x-μ i ||², where C i For the i-th cluster, μ i Using the cluster centers as the basis, plot the SSE-K curve, and select the K value corresponding to the inflection point of the curve as the optimal number of clusters; (2) Perform K-means clustering The K-means algorithm is used to cluster the variety feature matrix D: ① Randomly select K initial cluster centers μ1, μ2, ..., μ k ; ② Assign each variety to the nearest cluster center: C i ={X' j ||X' j -μ i ||≤||X'j -μ k ||,∀k≠i}; ③ Update cluster centers: μ i =(1 / |C i |)Σ X ∈C iX ; ④ Repeat steps ②-③ until the cluster centers no longer change or the maximum number of iterations is reached; (3) Calculate the clustering quality index The silhouette coefficient is used to evaluate the clustering quality: For variety i, its silhouette coefficient is calculated as: s(i)=(b(i)-a(i)) / max{a(i),b(i)}, where: a(i) is the average distance between variety i and other varieties in the same cluster; b(i) is the average distance between variety i and varieties in the nearest cluster; the silhouette coefficient ranges from [-1,1], and the larger the value, the better the clustering quality.

[0007] Preferably, step S3 determines the specific content of the cluster to which the current variety belongs as follows: (1) Extracting the characteristics of the current variety For the variety k to be produced, extract its variety feature vector X. k Standardization yields X' k ; (2) Calculate the distance to each cluster center Calculate X' k Euclidean distance to each cluster center: d i =||X' k -μ i ||; (3) Assign to the nearest cluster Assign the current variety to the nearest cluster C. m m = argmin(d1, d2, ..., d k ).

[0008] Preferably, the specific content of step S4, which extracts the historical quality data of homo-clustered varieties, is as follows: (1) Define quality characteristics For precision strip steel products, the key quality characteristics are defined as follows: Y1 is strip thickness (mm); Y2 is strip width (mm); Y3 is thickness deviation (%); Y4 is shape index (IU); Y5 is surface roughness (μm); Y6 is tensile strength (MPa); Y7 is yield strength (MPa); Y8 is elongation (%); Y9 is hardness (HV); Y 10 Flatness (mm / m); (2) Extract data of homo-clustered varieties Extracting clusters C from the quality database m Historical quality data for all varieties in China; (3) Abnormal data removal Outliers were removed using the 3σ criterion: the mean μ of the j-th quality characteristic of variety i was calculated. ij and standard deviation σ ij Remove data that meets the following conditions: |y ijk -μ ij |>3σ ij .

[0009] Preferably, the specific content of step S5, which involves data fusion based on similarity weighting, is as follows: (1) Calculate the variety similarity For cluster C m Given variety i, calculate its similarity to the current variety k: S i =exp(-||X' k -X' i ||²), similarity S i The value ranges from [0,1], with a larger value indicating a higher similarity. (2) Calculate the fusion weights The fusion weight of variety i is: w i =S i / Σ j ∈C m S j The weights satisfy: Σ i ∈C m w i =1; (3) Calculate the fusion statistic For the j-th quality characteristic, the fusion mean and standard deviation are calculated after similarity weighting: Fusion mean μ j *=Σ i ∈C m w i μ ij , fusion standard deviation σ j *=√(Σ i ∈C m w i ·σ ij 2 ).

[0010] Preferably, the specific content of step S6, adaptive control limit calculation, is as follows: (1) Initial control limits Based on the fusion statistics, calculate the initial control limits for the j-th quality characteristic: Centerline: CLo j =T j ; Upper control limit: UCLoo j =T j +3σ j *; Lower control limit: LCLo j =T j -3σ j *; where T j Let j be the target value of the j-th quality characteristic; (2) Adaptive weights As current variety data accumulates, the weights are dynamically adjusted: α t =min(t / 30,0.7), β t =1-α t ;wherein: α t β represents the weight of the current variety data, which increases with the sample size t, with an upper limit of 0.7; t To integrate data weights, the weights decrease as the number of samples increases, where t is the current number of samples; (3) Update control limits Calculate the statistical measure μ of the current variety. kj and σ kj Calculate the adaptive standard deviation: σ j adapt =√(α t ·σ kj 2 +β t ·σ j * 2 ); Update control limits: CL j =α t μ kj +β t ·T j UCL j =CL j +3σ j adapt LCL j =CL j -3σ j adapt , where: α t β represents the weight of the current variety data. t To integrate data weights.

[0011] Preferably, the specific content of step S7, multi-quality characteristic integrated control, is as follows: (1) Calculate T 2 Statistic For the i-th sample, calculate the standardized quality characteristic vector: Z i =[(yi1 -CL1) / σ1 adapt ,...,(y i10 -CL 10 ) / σ 10 adapt ] T , Calculate the Hotelling T² statistic: T 2 i =Z i T ·S -1 ·Z i Where S is the adaptive covariance matrix, S=α t ·S k +β t ·S * S * To integrate the covariance matrix, S k This is the covariance matrix of the current variety, calculated based on the accumulated samples of the varieties currently in production. (2) T 2 Control Limits UCL_T 2 =((t-1)p) / (tp)·F(p,tp,α)·(1+β t ·1.2), where p is the number of quality characteristics (p=10), t is the current number of samples, and F is the critical value of the F distribution; (3) Contribution analysis When T 2 When the statistic exceeds the control limit, calculate the contribution of each quality characteristic: C j =|Z ij ·(S -1 ·Z i ) j | / T 2 i The quality characteristic that contributes the most is the main source of anomalies.

[0012] Preferably, the specific content of step S8, abnormal pattern identification, is as follows: (1) Define the exception judgment rules Rule 1: Two out of three consecutive points fall outside the 2σ region on the same side of the center line; Rule 2: Four out of five consecutive points fall outside the 1σ region on the same side of the center line; Rule 3: Eight consecutive points fall on the same side of the center line; Rule 4: Six consecutive points showing a monotonous increase or decrease; Rule 5: Alternate up and down for 14 consecutive points; Rule 6: 15 consecutive points fall within the 1σ region on both sides of the center line; (2) Classification of abnormal patterns According to the anomaly detection rules and T 2 The trend of statistical changes categorizes outlier patterns into: Type 1: Sudden anomaly - single or a few point mutations, large amplitude, short duration; Type 2: Trend anomaly - continuous unidirectional change, gradual deviation from the center line; Type 3: Periodic anomalies - regular fluctuations that repeat at fixed intervals; Type 4: Systemic shift - The centerline shifts across the board, forming a new stable level; Type 5: Dispersion Anomaly - Changes in data dispersion, either too concentrated or too scattered; (3) Diagnosis of abnormal causes Establish a rule base for associating abnormal patterns with process parameters: .

[0013] Preferably, the specific content of the process capability assessment in step S9 is as follows: (1) Single quality characteristic process capability index Cp kj =min((USL j -μ kj ) / (3σ kj ),(μ kj -LSL j ) / (3σ kj )); When the sample size is small (t<30), the standard Cp kj To correct the impact of fused data on the results, the corrected process capability index Cp is used. kj *=Cp kj ·√(1-β t ·(σ j * / σ kj -1) 2 ); (2) Multi-quality characteristic integrated process capability index Cp k _multi=(Π j=1 10 Cp kj *)^(1 / 10); (3) Process capability level determination Cp k _multi≥1.67: Process capability is sufficient (Grade A); 1.33≤Cp k_multi<1.67: Process capability is acceptable (Grade B); 1.00≤Cp k _multi < 1.33: Insufficient process capability (Level C); Cp k _multi<1.00: Process capability is severely inadequate (Level D).

[0014] Preferably, the specific content of the dynamic update of the control chart in step S10 is as follows: (1) Clustering is updated regularly After each batch is completed, the quality data of that batch is added to the variety history database, and K-means clustering is re-executed monthly to update the cluster centers and variety groups. (2) Adaptive adjustment of fusion weights Based on the control chart's performance, dynamically adjust the fusion weight parameters: If the misclassification rate is greater than 5%, increase the weight β of the fused data. t : If the false negative rate is greater than 10%, increase the weight α of the current variety data. t : (3) Quality prediction and early warning Based on the current variety quality data series, the exponential smoothing method is used to predict future quality trends: , where α_smooth is the smoothing coefficient, usually taken as 0.2.

[0015] Based on the relationship between predicted values ​​and control limits, a three-level early warning system is established: Level 1 Warning (Red): The predicted value exceeds the control limit; immediately stop the machine for inspection. Level II Warning (Orange): The predicted value has entered the 2σ region; monitoring should be strengthened. Level 3 Warning (Yellow): The predicted value enters the 1σ region or shows a monotonic trend; pay attention to the trend.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: (1) By using K-means clustering analysis and data fusion technology, we can make full use of the historical quality data of varieties in the same cluster, and establish a reliable control chart for the first production of new varieties, effectively solving the problem of insufficient sample size for a single variety and avoiding insufficient data in small-batch production; (2) An adaptive control limit calculation method is adopted. As current variety data accumulates, the weights of the fused data and the current data are dynamically adjusted so that the control limits gradually approach the actual process characteristics of the current variety, thereby improving the accuracy and adaptability of the control chart. (3) Through HotellingT 2Statistical and contribution analysis enables comprehensive monitoring of 10 quality characteristics, timely detection of abnormal correlations between quality characteristics, and comprehensive monitoring of multiple quality characteristics. (4) Establish a cluster-based variety grouping mechanism, and achieve effective integration and utilization of different variety data through similarity weighting, thereby improving the utilization rate of historical data; (5) By identifying abnormal patterns and associating rule bases, the causes of abnormalities can be quickly located, thereby enhancing the ability to diagnose abnormalities; (6) By using quality trend prediction and multi-level early warning mechanism, early warning can be given before quality anomalies occur, realizing the transformation from passive control to active prevention, and achieving quality prediction and active early warning. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the embodiments.

[0018] Example 1: Production process control of 304 stainless steel precision strip A precision strip steel manufacturer produces 304 stainless steel precision strip steel with a specification of 0.15mm × 200mm, with an annual output of approximately 500 tons, produced in 20 batches, each batch weighing approximately 25 tons. The statistical process control method for small-batch, multi-variety precision strip steel, as described in this invention, is used for quality control during the production process. The specific steps are as follows: Step 1: Establish a variety characteristic database This company has historically produced 150 types of stainless steel precision strip products. Characteristic parameters for each type were extracted to establish a product feature database. For the 304 stainless steel precision strip (product number K001) currently to be produced, its feature vector is: X_K001=[0.15,200,304,520,205,40,0.4,0.005,180,2,1,1] The eigenvectors are standardized to obtain the standardized eigenvectors X'_K001.

[0019] Step 2: Variety grouping based on K-means clustering Using the elbow rule to calculate the SSE corresponding to different K values, it was found that the curve showed a clear inflection point when K=7, thus determining the optimal number of clusters K=7. K-means clustering was performed on 150 historical varieties, resulting in 7 clusters: Cluster C1: Ultra-thin foil materials, 18 varieties; Cluster C2: Thin-sized soft-state type, 25 varieties; Cluster C3: Medium-sized soft-state group, 22 varieties; Cluster C4: Medium-sized hard-state class, 28 varieties; Cluster C5: Thick-sized hard-state type, 20 varieties; Cluster C6: Extremely hard, 15 varieties; Cluster C7: Special surface type, 22 varieties; The overall average profile coefficient is 0.70, indicating good clustering quality.

[0020] Step 3: Determine the cluster to which the current variety belongs. Calculate the Euclidean distance from X'_K001 to the 7 cluster centers: d1=2.85, d2=0.68, d3=1.12, d4=1.95, d5=2.45, d6=3.20, d7=2.10. The smallest distance is in cluster C2. Therefore, variety K001 is assigned to cluster C2 (thin-size soft state).

[0021] Step 4: Extract quality data of homo-clustered varieties Historical quality data of 25 varieties in cluster C2 were extracted from the quality database, totaling 3,500 sample points; Outliers were removed using the 3σ criterion for each of the 10 quality characteristics, with a removal rate of approximately 2.5%, resulting in 3412 valid samples.

[0022] Step 5: Similarity-weighted data fusion Calculate the similarity between variety K001 and the 25 varieties in cluster C2: S1=0.052,S2=0.085,S3=0.125,S4=0.098,S5=0.142,S6=0.078,S7=0.095,S8=0.112,S9=0.068,S 10 =0.088,S 11 =0.105,S 12 =0.092,S 13 =0.118,S 14 =0.075,S 15 =0.082,S 16 =0.095,S 17 =0.108,S 18 =0.072,S 19 =0.088,S 20 =0.095,S 21 =0.102,S 22 =0.085,S 23 =0.078,S 24 =0.092,S 25 =0.088, normalization yields the fusion weight w1-w 25 .

[0023] The fusion statistics for the 10 quality characteristics are calculated as shown in the table below: .

[0024] Step 6: Adaptive control limit calculation Based on the fusion statistics, the initial control limits are calculated as shown in the table below: .

[0025] As samples of variety K001 accumulate, the control limits are dynamically updated: When t=10, α 10 =0.33, β 10 =0.67:

[0026] When t=20, α 20 =0.67, β 20 =0.33:

[0027] When t=30, α 30 =0.70, β 30 =0.30, control limits are stable.

[0028] Step 7: Integrated Control of Multiple Quality Characteristics Calculate the covariance matrix S, and calculate T for each sample. 2 Statistics.

[0029] Calculate T 2 Control limit: UCL_T 2 =23.5 During the production process, the T of the 15th sample 2 The statistic is 25.8, which exceeds the control limit.

[0030] Calculate contribution: C1=0.08,C2=0.05,C3=0.12,C4=0.35,C5=0.08,C6=0.06,C7=0.05,C8=0.07,C9=0.09,C 10 =0.05, with Y4 (plate shape) contributing the most, making it the main source of anomalies.

[0031] Step 8: Anomaly Pattern Recognition Analysis of the plate shape data of samples 10-20 revealed a continuous monotonous increase of 6 points, which conforms to rule 4 and is judged as a trend anomaly.

[0032] The query for the association rule base may fail due to the following reasons: Roll wear, tension system drift, or unstable temperature control.

[0033] Diagnostic recommendations: Check the wear of the rolls, calibrate the tension control system, and check the accuracy of the temperature control.

[0034] Step 9: Process Capability Assessment Calculation The process capability indices for each quality characteristic are shown in the table below:

[0035] Comprehensive process capability index: Cp k _multi=(1.15×1.20×3.20×0.72×0.59×1.13×0.94×1.25×1.33×1.53)^(1 / 10)=1.18, the process capability level is C (insufficient process capability), mainly affected by Y4 (plate shape) and Y5 (roughness).

[0036] Step 10: Dynamically update the control chart Predicting plate shape trend using exponential smoothing: =0.2×3.15+0.8×3.08=3.09,The predicted value enters the 2σ region (2.55-5.10), triggering a level II warning (orange).

[0037] The system sends warning messages to operators and process engineers, suggesting that they strengthen monitoring and check the rolls and tension system.

[0038] After adopting the method of this invention, the quality control effect of product K001 has been significantly improved. The overall pass rate has increased from 94.2% to 98.5%, the board shape pass rate has increased from 92.5% to 97.8%, the non-conforming rate has decreased from 3.8% to 1.2%, and the number of customer complaints has decreased from 5 times / year to 1 time / year.

[0039] Example 2: Production process control of 316L ultrathin foil A precision strip steel manufacturer produces 316L stainless steel ultra-thin foil with a specification of 0.03mm × 100mm. The annual output is approximately 80 tons, produced in 8 batches, with each batch weighing approximately 10 tons. This is a new product, and there is no historical data available.

[0040] Steps 1-3: Variety Feature Extraction and Cluster Assignment Extracting the feature vector of variety K178: X_K178=[0.03,100,316L,580,240,45,0.2,0.002,165,4,1,6] After standardization, the distance to each cluster center is calculated and assigned to cluster C1 (ultra-thin foil).

[0041] Steps 4-5: Data Fusion Cluster C1 contains 18 historical varieties, and 2160 sample points of historical quality data were extracted.

[0042] The similarity and fusion weights were calculated, and the fusion statistics are shown in the table below:

[0043] Step 6: Adaptive Control Limits Since it is a new variety, the initial data relied entirely on fused data (α0=0, β0=1).

[0044] As the sample size accumulates, the control limits are gradually adjusted: When t=5: α5=0.17, β5=0.83; At t=10: α 10 =0.33, β 10 =0.67; At t=15: α 15 =0.50, β 15 =0.50; Steps 7-9: Quality Monitoring and Capability Assessment During the production process, T 2 No abnormalities were found in the control charts.

[0045] When the sample size reaches 30, the computational capability is: Cp k _multi=1.42, process capability level is B (process capability is acceptable).

[0046] After adopting the method of this invention, an effective control chart was established in the first production of the new variety, with an overall pass rate of 96.8%, no quality abnormalities occurred, and customer satisfaction was high.

[0047] Example 3: Production process control of 430 special surface strip steel A precision strip steel manufacturer produces 430 stainless steel special surface strip with specifications of 0.25mm × 300mm, with an annual output of approximately 200 tons, produced in 12 batches. This product requires a surface roughness Ra ≤ 0.15μm, exhibiting extremely high quality requirements.

[0048] Steps 1-5: Variety Grouping and Data Fusion Variety K052 was assigned to cluster C7 (Special Surfaces), which contains 22 varieties; Due to the special requirements for surface quality, Y5 (roughness) is given a higher weight.

[0049] Steps 6-7: Adaptive Control and Integrated Monitoring During the production of the 8th batch, the 12th sample T 2 The statistical limit has been exceeded.

[0050] Contribution analysis showed that Y5 (roughness) contributed 0.42, making it the main source of anomalies.

[0051] Step 8: Abnormal Diagnosis Analysis of the roughness data revealed that 4 out of 5 consecutive points fell outside the 1σ region and exhibited periodic fluctuations, which was identified as a periodic anomaly.

[0052] Possible reasons for querying the association rule base: Problems with the surface quality of the rolls, fluctuations in rolling speed, or unstable viscosity of the rolling oil; Diagnostic recommendations: Inspect the surface quality of the rolls, replace the rolls if necessary, check the speed control system, and check the rolling oil viscosity control. Upon inspection, minor scratches were found on the surface of the rolls. The problem was resolved after replacing the rolls.

[0053] Steps 9-10: Capacity Assessment and Early Warning After replacing the rolls, reassess the process capability: Cp k _multi=1.58 (up from 1.12), process capability level improved from C to B.

[0054] Establish a quality prediction model to predict and warn of roughness in real time, effectively avoiding similar problems in the future.

[0055] After implementing the method of this invention, the proportion of varieties with control charts can be increased from 35% to 100%. The average sample requirement was reduced from 25-30 to 10-15; the false positive rate decreased from 0.85% to 0.49%, a reduction of 42%; the false negative rate decreased from 12.5% ​​to 6.7%, a reduction of 46%; the control chart creation time was shortened from 8-12 hours to 2-4 hours, a reduction of 67%; the anomaly detection rate increased by more than 25%; it can identify associated anomalies that cannot be detected by univariate control charts; the number of control charts was reduced from more than 10 to 1 integrated control chart; the data utilization rate increased from 30% to more than 90%, an increase of 60%; the historical data accumulation period was shortened from 6-12 months to 1-2 months; historical data of 165 varieties were fully utilized; the anomaly diagnosis accuracy rate increased from 55% to 82%, an increase of 49%; the anomaly processing time was shortened from 2-4 hours to 1-2 hours, a reduction of 50%; and an identification rule and diagnostic knowledge base containing 5 anomaly types were established.

Claims

1. A statistical process control method for precision strip steel with small batch and multiple varieties, characterized in that, The specific steps are as follows: Step S1: Establish a database of precision strip steel product characteristics Extract multidimensional feature vectors and perform Z-score standardization; Step S2: Group historical varieties based on K-means clustering algorithm The elbow rule is used to determine the optimal number of clusters, and the silhouette coefficient is used to evaluate the cluster quality. Step S3: Determine the cluster to which the current variety belongs. Calculate the Euclidean distance of the current variety to each cluster center, and assign it to the cluster with the closest distance; Step S4: Extract historical quality data of homo-clustered varieties and remove outliers using the 3σ criterion. Step S5: Similarity-weighted data fusion Variety similarity is calculated based on Euclidean distance, and data fusion is performed using similarity weighting to calculate fusion statistics; Step S6: Adaptive control limit calculation The weights of fused data and current variety data are dynamically adjusted as samples accumulate. Step S7: Integrated Control of Multiple Quality Characteristics The Hotelling T² statistic is used to achieve comprehensive monitoring of multiple quality characteristics, and the main sources of anomalies are identified through contribution analysis. Step S8: Abnormal Pattern Recognition Establish an anomaly pattern recognition mechanism and association rule base to achieve intelligent diagnosis of anomaly causes; Step S9: Process Capability Assessment Calculate the comprehensive process capability index of multiple quality characteristics to achieve accurate assessment of process capability; Step S10: Dynamically update the control chart Establish a dynamic update mechanism for control charts, including periodic updates of clustering, adaptive adjustment of fusion weights, and quality prediction and early warning.

2. The statistical process control method for precision strip steel in small-batch, multi-variety production according to claim 1, characterized in that, The specific details of step S1, establishing the precision strip steel product feature database, are as follows: (1) Define the variety feature vector For each precision strip steel grade, feature parameters are extracted to form a grade feature vector X: X=[x1,x2,x3,x4,x5,x6,x7,x8,x9,x 10 x 11 x 12 [The following is a list of parameters:] where: x1 is thickness, ranging from 0.02 to 0.5 mm; x2 is width, ranging from 300 to 650 mm; x3 is material type (AISI 200, 300, 400 series); x4 is tensile strength, ranging from 300 to 2100 MPa; x5 is yield strength, ranging from 170 to 2000 MPa; x6 is elongation, ranging from 2% to 65%; x7 is surface roughness, Ra value 0.1 to 0.5 μm; x8 is thickness tolerance, ranging from ±0.001 to ±0.01 mm; x9 is hardness, ranging from 150 to 605 HV; 10 The number of rolling passes, ranging from 1 to 4; x 11 Annealing process type (bright annealing, bell-type furnace annealing, Ta annealing); x 12 For product series (soft state, hard state, extremely hard, Ta, special surface, ultra-thin foil); (2) Feature standardization Z-score normalization is applied to the variety feature vector: x' i =(x i -μ i ) / σ i , where μ i and σ i and are the mean and standard deviation of the i-th feature across all varieties, respectively.

3. The statistical process control method for precision strip steel with small batch and multiple varieties according to claim 2, characterized in that, The specific details of grouping historical varieties in step S2 are as follows: (1) Determine the number of clusters The Elbow Method is used to determine the optimal number of clusters K, and the within-group sum of squares SSE corresponding to different K values ​​(K=3,4,5,...,15) is calculated as follows: SSE(K)=Σ i=1 K Σ X ∈C i ||x-μ i ||², where C i For the i-th cluster, μ i Using the cluster centers as the basis, plot the SSE-K curve, and select the K value corresponding to the inflection point of the curve as the optimal number of clusters; (2) Perform K-means clustering The K-means algorithm is used to cluster the variety feature matrix D: ① Randomly select K initial cluster centers μ1, μ2, ..., μ k ; ② Assign each variety to the nearest cluster center: C i ={X' j ||X' j -μ i ||≤||X' j -μ k ||,∀k≠i}; ③ Update cluster centers: μ i =(1 / |C i |)Σ X ∈C iX ; ④ Repeat steps ②-③ until the cluster centers no longer change or the maximum number of iterations is reached; (3) Calculate the clustering quality index Cluster quality is assessed using the silhouette coefficient: For variety i, its silhouette coefficient is calculated as follows: s(i)=(b(i)-a(i)) / max{a(i),b(i)}, where: a(i) is the average distance between variety i and other varieties in the same cluster; b(i) is the average distance between variety i and varieties in the nearest cluster; the silhouette coefficient ranges from [-1,1], and the larger the value, the better the clustering quality.

4. The statistical process control method for precision strip steel with small batch and multiple varieties according to claim 3, characterized in that, The specific details of determining the cluster to which the current variety belongs in step S3 are as follows: (1) Extracting the characteristics of the current variety For the variety k to be produced, extract its variety feature vector X. k Standardization yields X' k ; (2) Calculate the distance to each cluster center Calculate X' k Euclidean distance to each cluster center: d i =||X' k -μ i ||; (3) Assign to the nearest cluster Assign the current variety to the nearest cluster C. m m = argmin(d1, d2, ..., d k ).

5. The statistical process control method for precision strip steel with small batch and multiple varieties according to claim 4, characterized in that, The specific details of step S4, which extracts historical quality data of homo-clustered varieties, are as follows: (1) Define quality characteristics For precision strip steel products, the key quality characteristics are defined as follows: Y1 is strip thickness (mm); Y2 is strip width (mm); Y3 is thickness deviation (%); Y4 is shape index (IU); Y5 is surface roughness (μm); Y6 is tensile strength (MPa); Y7 is yield strength (MPa); Y8 is elongation (%); Y9 is hardness (HV); Y 10 Flatness (mm / m); (2) Extract data of homo-clustered varieties Extracting clusters C from the quality database m Historical quality data for all varieties in China; (3) Abnormal data removal Outliers were removed using the 3σ criterion: the mean μ of the j-th quality characteristic of variety i was calculated. ij and standard deviation σ ij Remove data that meets the following conditions: |y ijk -μ ij |>3σ ij .

6. The statistical process control method for precision strip steel with small batch and multiple varieties according to claim 5, characterized in that, The specific details of step S5, which involves data fusion based on similarity weighting, are as follows: (1) Calculate the variety similarity For cluster C m Given variety i, calculate its similarity to the current variety k: S i =exp(-||X' k -X' i ||²), similarity S i The value ranges from [0,1], with a larger value indicating a higher similarity. (2) Calculate the fusion weights The fusion weight of variety i is: w i =S i / Σ j ∈C m S j The weights satisfy: Σ i ∈C m w i =1; (3) Calculate the fusion statistic For the j-th quality characteristic, the fusion mean and standard deviation are calculated after similarity weighting: Fusion mean μ j *=Σ i ∈C m w i μ ij , fusion standard deviation σ j *=√(Σ i ∈C m w i ·σ ij 2 ).

7. The statistical process control method for precision strip steel with small batch and multiple varieties according to claim 6, characterized in that, The specific details of step S6, adaptive control limit calculation, are as follows: (1) Initial control limits Based on the fusion statistics, calculate the initial control limits for the j-th quality characteristic: Centerline: CLo j =T j ; Upper control limit: UCLoo j =T j +3σ j *; Lower control limit: LCLo j =T j -3σ j *; where T j Let j be the target value of the j-th quality characteristic; (2) Adaptive weights As current variety data accumulates, the weights are dynamically adjusted: α t =min(t / 30,0.7), β t =1-α t ;wherein: α t β represents the weight of the current variety data, which increases with the sample size t, with an upper limit of 0.7; t To integrate data weights, the weights decrease as the number of samples increases, where t is the current number of samples; (3) Update control limits Calculate the statistical measure μ of the current variety. kj and σ kj Calculate the adaptive standard deviation: s j adapt =√(a t ·s kj 2 +b t ·s j * 2 ); Update control limits: CL j =α t μ kj +β t ·T j UCL j =CL j +3σ j adapt LCL j =CL j -3σ j adapt , where: α t β represents the weight of the current variety data. t To integrate data weights.

8. The statistical process control method for precision strip steel with small batch and multiple varieties according to claim 7, characterized in that, The specific details of step S7, multi-quality characteristic integrated control, are as follows: (1) Calculate T 2 Statistic For the i-th sample, calculate the standardized quality characteristic vector: Z i =[(y i1 -CL1) / σ1 adapt ,...,(y i10 -CL 10 ) / σ 10 adapt ] T , Calculate the Hotelling T² statistic: T 2 i =Z i T ·S -1 ·Z i Where S is the adaptive covariance matrix, S=α t ·S k +β t ·S * S * To integrate the covariance matrix, S k This is the covariance matrix of the current variety, calculated based on the accumulated samples of the varieties currently in production. (2) T 2 Control Limits UCL_T 2 =((t-1)p) / (tp)·F(p,tp,α)·(1+β t ·1.2), where p is the number of quality characteristics (p=10), t is the current number of samples, and F is the critical value of the F distribution; (3) Contribution analysis When T 2 When the statistic exceeds the control limit, calculate the contribution of each quality characteristic: C j =|Z ij ·(S -1 ·Z i ) j | / T 2 i The quality characteristic that contributes the most is the main source of anomalies.

9. A statistical process control method for precision strip steel in small-batch, multi-variety production according to claim 8, characterized in that, The specific details of step S8, abnormal pattern identification, are as follows: (1) Define the exception judgment rules Rule 1: Two out of three consecutive points fall outside the 2σ region on the same side of the center line; Rule 2: Four out of five consecutive points fall outside the 1σ region on the same side of the center line; Rule 3: Eight consecutive points fall on the same side of the center line; Rule 4: Six consecutive points showing a monotonous increase or decrease; Rule 5: Alternate up and down for 14 consecutive points; Rule 6: 15 consecutive points fall within the 1σ region on both sides of the center line; (2) Classification of abnormal patterns According to the anomaly detection rules and T 2 The trend of statistical changes categorizes outlier patterns into: Type 1: Sudden anomaly - single or a few point mutations, large amplitude, short duration; Type 2: Trend anomaly - continuous unidirectional change, gradual deviation from the center line; Type 3: Periodic anomalies - regular fluctuations that repeat at fixed intervals; Type 4: Systemic shift - The centerline shifts across the board, forming a new stable level; Type 5: Dispersion Anomaly - Changes in data dispersion, either too concentrated or too scattered; (3) Diagnosis of abnormal causes Establish a rule base for associating abnormal patterns with process parameters:

10. A statistical process control method for precision strip steel in small-batch, multi-variety production according to claim 9, characterized in that, The specific content of the process capability assessment in step S9 is as follows: (1) Single quality characteristic process capability index Cp kj =min((USL j -m kj ) / (3σ kj ),(m kj -LSL j ) / (3σ kj )); When the sample size is small (t<30), the standard Cp kj To correct the impact of fused data on the results, the corrected process capability index Cp is used. kj *=Cp kj ·√(1-β t ·(σ j * / σ kj -1) 2 ); (2) Multi-quality characteristic integrated process capability index Cp k _multi=(Π j=1 10 Cp kj *)^(1 / 10); (3) Process capability level determination Cp k _multi≥1.67: Process capability is sufficient (Grade A); 1.33≤Cp k _multi<1.67: Process capability is acceptable (Grade B); 1.00≤Cp k _multi < 1.33: Insufficient process capability (Level C); Cp k _multi < 1.00: Process capability is severely inadequate (Grade D); The specific details of step S10, dynamic updating of the control chart, are as follows: (1) Clustering is updated regularly After each batch is completed, the quality data of that batch is added to the variety history database, and K-means clustering is re-executed monthly to update the cluster centers and variety groups. (2) Adaptive adjustment of fusion weights Based on the control chart's performance, dynamically adjust the fusion weight parameters: If the misclassification rate is greater than 5%, increase the weight β of the fused data. t : If the false negative rate is greater than 10%, increase the weight α of the current variety data. t : (3) Quality prediction and early warning Based on the current variety quality data series, the exponential smoothing method is used to predict future quality trends: , where α_smooth is the smoothing coefficient, usually taken as 0.2; Based on the relationship between predicted values ​​and control limits, a three-level early warning system is established: Level 1 Warning (Red): The predicted value exceeds the control limit; immediately stop the machine for inspection. Level II Warning (Orange): The predicted value has entered the 2σ region; monitoring should be strengthened. Level 3 Warning (Yellow): The predicted value enters the 1σ region or shows a monotonic trend; pay attention to the trend.

Images